Defect detection in polymer components using infrared thermography by employment of signal processing methods and U-Net semantic segmentation

In many industrial applications, accurate defect detection in polymer components is crucial to ensure safety and reliability. This study addresses the limited research in this area by developing and comparing active thermographic processing techniques coupled with a deep learning U-Net model for automatic detection of subsurface defects in HDPE, PTFE, and PCTFE polymeric materials. For this purpose, a dataset including samples with flat bottom holes of different dimensions and depths was prepared and the image process was performed during the cooling stage. Thermal data processing was performed with advanced techniques such as PCA, ICA, TSR, PLSR, TPPT, PICA and CLAHE to improve the signal-to-noise ratio and image quality. In the segmentation stage, the U-Net model with MobileNetV2 as the backbone was used and the accuracy was 96.55%, recall 85.96%, precision 93.84%, dice coefficient of 89.38% and IoU of 80.97% were obtained. Also, the processing time for each image was measured to be only 0.0952 seconds. The best results were obtained from the PCA+UNet combination, which succeeded in identifying 100% POD defects (25 out of 25) with an SNR of 23.645 dB, ultimately making this method a fast and reliable option for real industrial applications.